Deep Advertising Conversion Optimization at Ximalaya

Preface

In the evolution of online advertising, customers have become increasingly strict about the value of ads. Their original intention is to obtain higher advertising revenue, i.e., larger ad spend and higher ROI. The deeper the ad‑chain optimization, the more accurate the traffic distribution, and the more guaranteed the ROI, which in turn raises the ceiling for budgets and bids. Therefore, the evolution of billing models has gradually moved from shallow to deep.

Online Advertising Evolution

For the platform: The closer the selling model is to the backend, the longer the optimization chain, the greater the revenue risk, and the more technical challenges arise.

Example: If client A sets a very high conversion bid, after a stable period of OCPC the conversion data may fluctuate sharply (due to chain issues or intentional data loss). This can cause the platform’s CVR estimate to be overly optimistic, and during the re‑convergence of the model and bid parameters the platform must bear the loss caused by the over‑estimation according to its compensation rules.

For advertisers: The closer the selling model is to the backend, the more controllable the risk and the higher the profit space, which motivates them to increase budget and bid.

Example: Client B originally used CPC, receiving high‑click traffic without guaranteed downstream conversion. By switching to OCPC with a set conversion bid (e.g., cost per added WeChat, cost per download), the conversion rate improves and the advertiser only bears the conversion cost → ROI fluctuation, while the exposure‑to‑conversion risk is transferred to the platform.

Advertising Chain Optimization Path

The design of advertising algorithm strategies aims to solve real business difficulties, improve user experience, and increase traffic distribution efficiency. Ximalaya’s ad‑delivery chain optimization is a long‑term, continuous effort. Over the past year, we have focused on a win‑win goal for the platform and advertisers, striving to improve deep‑level conversion performance. The following sections describe our exploration process.

Ximalaya Advertising Deep Conversion Optimization

In the past year, deep‑level conversion optimization has proceeded in three stages: OCPC deep dual‑bidding, full‑channel model assistance, and real‑time crowd premium.

The three deep‑optimization modes complement each other. Their differences lie in conversion delay and data volume, allowing deep optimization to cover more client types (new/old, long/short data cycles).

(1) The three modes are mutually complementary. Their distinction mainly concerns conversion latency and data scale, enabling deep‑level effect optimization to serve a broader range of customers.

(2) Deep OCPC is the main tool for improving backend conversion, but for clients with very long conversion delays the platform’s compensation risk becomes too high, so we introduce full‑channel assistance and real‑time crowd premium as supplements.

(3) Full‑channel assistance and real‑time crowd premium are progressive: the former expands the audience using off‑site data when internal data are insufficient, while the latter, once internal data are abundant, switches to fine‑grained traffic distribution for better backend results.

Technical Challenges

The ad‑delivery chain can be simplified as: exposure → click → shallow conversion (form, activation) → deep conversion (retention, credit, payment) → ROI (payment amount, credit limit). Optimizing ROI means gradually moving the platform’s predictive model and ranking target from the bidding point to the advertiser’s evaluation point.

However, as the chain lengthens, optimization difficulty increases. The main problems are:

(1) Data sparsity: Each additional layer reduces data volume by orders of magnitude, making sample accumulation difficult and causing large fluctuations in single‑conversion metrics.

(2) Extreme conversion delay: Deep conversion goals such as 7‑day retention or payment can be delayed by 1–2 weeks, leading to temporal bias in model samples.

(3) Cold‑start for new clients: Shallow conversion data accumulate relatively quickly, but deep conversion data may require months, creating uncertainty for both platform and advertiser.

(4) Unstable traffic & diverse conversion types: Traffic volume fluctuates across time, and conversion rates vary widely among the >20 conversion goals, complicating joint modeling for ranking correctness and estimation precision.

OCPC Deep Dual‑Bidding

Core goal: Optimize advertisers’ backend conversion and guarantee backend conversion cost.

Advertisers seek higher ROI; when a client has better backend optimization tools, its bid ceiling rises, breaking the previous equilibrium and capturing more traffic. Media platforms also compete for limited advertiser budgets, continuously improving product and traffic distribution efficiency.

(1) Business pain points

Platform revenue bottleneck: When trying to raise client bids, they often reply that they are already at cost and will only increase bids if ROI improves.

Insufficient traffic competition: High‑ROI traffic is contested by many clients, while lower‑quality traffic is undervalued.

Poor client experience: Direct OCPC deep targeting leads to unstable cost and performance.

Understanding users: Optimizing backend conversion requires access to downstream data (payments, adds, long‑term retention).

(2) Solutions

Deep CVR optimization: Instead of letting advertisers manually adjust targeting, the platform intervenes with deep models for fine‑grained optimization.

Dual‑bidding: Use a shallow bid for cold‑start traffic while data accumulate, then automatically transition to deep‑bid once sufficient data are available.

Goal: Gradually shift user groups from shallow to deep conversion, raising profit space, budget, and bid ceilings.

(3) Technical approach

Conversion‑delay data correction: Define confidence intervals to ensure modules use stable, high‑confidence data.

Deep conversion modeling: Merge shallow and deep CVR models, combine multiple business lines, and use multi‑task learning to let shallow CVR assist embedding learning.

Deep‑shallow weight transfer: When deep data are scarce, shallow eCPM weight assists ranking; as deep data grow, increase deep weight.

(4) Client cases

Client

Delivery chain

User characteristics

OCPC deep dual‑bidding effect (vs. OCPC)

Private‑domain lead client

Play → Add WeChat (delay 1‑7 days)

Very low deep conversion rate, weak chain correlation

+200% add‑WeChat rate

Financial client

Form → Credit (delay 1‑7 days)

ROI‑sensitive; poor ROI leads to budget cut

+27% credit approval rate

Full‑Channel Data Assistance

Core goal: When conversion delay is large, improve backend conversion using off‑site data and solve internal cold‑start problems.

After OCPC deep dual‑bidding, ROI improves for many clients, but coverage plateaus. Some clients experience 1‑2 week conversion delays, causing compensation risk and unstable volume.

(1) Business pain points

Data islands: Cross‑media campaigns create isolated conversion data, limiting optimization.

Long cold‑start cycles: Even with dual‑bidding, lack of data can take weeks, during which clients may churn.

Large back‑transfer delay: Platform compensation rules cannot cover very long delays fairly.

Uneven data back‑transfer capability: Some clients can programmatically send data; others rely on manual sales follow‑up.

(2) Solutions

External data introduction: Use off‑site conversion behavior to accelerate model convergence for new clients.

Audience recall: Treat high‑quality off‑site converters as seed users and build look‑alike models to expand the audience.

Multiple back‑transfer methods: Support both API integration and periodic file uploads.

(3) Modeling ideas

Model construction: Require clients to label behavior types (payment, add‑WeChat, etc.) for more precise modeling.

Feature selection: Add ad‑related IDs, client categories, and behavior types to the existing user/device profile.

Sample selection: In addition to random negative sampling, add hard samples based on historical records to improve discrimination between “good front‑end / poor back‑end” users.

(4) Example clients

Client

Delivery chain

User characteristics

Full‑channel assistance effect (vs. OCPC)

Broadcast education client

Form → Add WeChat → Payment (delay 1‑2 weeks)

Very long conversion delay

+28% course purchase rate, +32% ROI

Adult English client

Form → Credit (delay 1‑2 weeks)

Very long conversion delay, lack of internal data

+27% credit approval rate

Real‑Time Crowd Premium

Core goal: When conversion delay is large, push the optimization capability to the extreme by applying fine‑grained traffic‑level ROI weighting.

Full‑channel assistance works well for cold‑start, but once internal data are sufficient, continuing to rely on audience recall becomes inefficient. Real‑time crowd premium optimizes traffic at the granularity of user × creative × context.

(1) Business pain points

Off‑site paid users may not be willing to pay on the current platform.

Audience recall ignores important features such as ad placement, creative, landing page, and real‑time behavior, limiting its effectiveness.

(2) Solutions

Real‑time crowd premium modeling: Build a dedicated model for clients with extreme conversion delay.

Reasonable premium: Instead of directly multiplying the model’s estimate into ranking, map it to ROI improvement.

Seamless switch: Use full‑channel assistance for cold‑start, then transition to real‑time premium once internal data are sufficient.

(3) Modeling ideas

Model design: Class‑ESMM multi‑task structure, with CTR assisting CVR learning; treat deep conversion delay as a separate task.

Strategy design: Identify high‑ROI traffic and assign it a higher eCPM weight, while suppressing low‑ROI traffic, achieving ROI uplift without sacrificing overall volume.

(4) Example clients

Client

Delivery chain

User characteristics

Real‑time premium effect (vs. OCPC)

Broadcast education client

Form → Add WeChat → Payment (delay 1‑2 weeks)

Very long conversion delay

+28% course purchase rate, +32% ROI

Adult English client

Form → Credit (delay 1‑2 weeks)

Very long conversion delay, lack of internal data

+27% credit approval rate

Future Development

Although many attempts have improved deep conversion rates and ROI, further directions remain:

Joint modeling of conversion delay and CVR to correct sample bias.

Lowering the data‑volume threshold so that more clients can use the system.

Direct ROI optimization, eventually achieving ROI‑protected bidding.